trollkotze's picture
More explanation in patches/README.
c54c481 verified
|
raw
history blame
636 Bytes

Assumption is that you have llama.cpp and repeng directories under the same parent directory, and that Miqu is in llama.cpp/models/miqu-1-70b.q5_K_M.gguf.

  • Copy files from patches/repeng into your repeng diretory.
  • Apply the diff in patches/llama inside your llama.cpp directory.
  • Place trainvector.sh in their common parent directory.
  • Run ./trainvector.sh {name-of-your-vector} to generate a vector based on the prompts generated from your edited repeng/emotion_prompts.py

Or build a better pipeline.

Parameters for context and batch size may not be optimal, and can be too small if you have longer prompts. Adjust as needed.